176 trained · 71,377 answers marked · 100% of the evidence found
I build the things I needed, and I can lead a team to build yours.
A serial builder, looking for a next role — or the backing to launch what I have already built. Ten products and nine published libraries, eight of which you can use or run today. This year alone I have taken them from a problem to something running — and every number on this site is one you could check.
Voice as a vital sign. A patient records five short tasks a day in their own language, and Voxtar measures tremor, breathiness and timing against that person’s own baseline — then flags the drift.
Best oral presentation — West African College of Physicians, July 2026, University of Ibadan, before a panel of consultants.The numbers
HealthcareIn design02
AuraHealth
Voice triage on the queue, with escrowed payment, so an emergency admission is not held up at a cashier’s desk. Designed, not yet built — the page says so.
The missing proof layer for YouTube. The best teaching is already free — what is missing is evidence you can do the work at the end of it.
Graded labs, not a quiz: code that passes hidden tests, a model that reconciles, a deck that survives checks.The numbers
EducationLive04
Hypatia
Learning by domain isomorphism. Bring what you already know, and it maps that structure onto what you want to learn — then teaches you exactly where the mapping breaks.
A daily drill bot for MB BS Part III. Telegram quiz polls each morning, marked automatically, with a leaderboard at close of play — because solo spaced repetition is a solved problem that almost nobody sustains alone.
155 students · 71,377 limbs marked at 75.47% · ₦47,900 from 25 payers.The numbers
AIApache 2.006
Vectorless
Document retrieval with no chunking, no embeddings and no vector database. A PDF has no headings — only glyphs with a size and a position — so Vectorless rebuilds the structure from the page geometry and navigates to the exact section, tables included.
On FinanceBench: all evidence found, and the section it points at went from 183 pages to 6.The numbers
AIMIT07
NotebookLM, self-hosted
The parts of NotebookLM I actually use, on my own keys and my own machine. One Hono app serves the web build and the desktop build, so there is a single server to reason about rather than two.
MIT, and the whole of it runs locally — Postgres compiled into the app rather than a database to install.The numbers
AIMIT08
BridgeHook
Webhook observability with nothing to install. A public URL forwards Stripe or GitHub straight to your localhost, because the browser tab itself is the tunnel agent — and every event is captured, replayable and signature-checked.
The stream is held in a Cloudflare Durable Object, which is what makes leaving it open all day cost nothing.The numbers
AIIn build09
Mercala
A storefront built so an agent can shop it, not just a person — Java 21 and Spring Boot across four modules, with the catalogue, the cart and the conversation behind one set of contracts.
Hybrid search with no Elasticsearch — full-text and vectors reconciled inside Postgres.The numbers
AIBuilt10
Tether
Lend a device without losing it. Custody moves between two people by compare-and-swap, so two phones can never both believe they hold the same thing.
The API has zero npm dependencies — Bun’s own runtime is the whole server, and it compiles to one executable.The numbers
How we work
We do not guess what matters in a system. We measure it, then publish what
the measurement said — including the parts that went against us.
02The numbers we haveFinanceBench, 21 filings, cold cache. From our own evaluation docs.
What we can prove today.
Our headline number today is a benchmark,
not an install count.
Vectorless — generative extraction replaced by a batched judge
Measure
Before
After
Evidence-page coverage
0.800
1.000
Cost per document
$0.017–0.095
$0.0003–0.0015
Time per document
103–840 s
12–143 s
Generative calls per document
1
0
Median span holding a gold page
183 pages
36 pages
When adoption is the number worth showing, it will be the one here.
03What we doThree fields we work in, and the layer all three stand on.
Three fields.
One layer underneath.
01
Healthcare. Voxtar leads, with an IRB application at UCH for a
post-thyroidectomy voice cohort as the current priority. Cervical screening,
oncology pharmacology and HIV-resistance genomics run behind it. Every clinical
product plugs into one shared record core instead of rebuilding it.
02
Research. Systematic reviews, mechanistic modelling, and checking that a
model found real signal rather than a shortcut in the data. Nothing ships from
this pillar — it is the standard the other two are held to, applied
whether the output is a manuscript or a product.
03
Education. A cohort of 176 came through three evenings in August.
Coursified turns open video into graded courses, Hypatia teaches a new field
through one you already know, and MB3 Prepbot drills a medical finals syllabus
every morning.
04The layer we ownWritten for our own products, then released on their own.
Three libraries.
Ours, and open.
01
Vectorless — Apache 2.0. Document retrieval with no chunking, no
embeddings and no vector database. It rebuilds a document’s structure from the geometry of the
page and navigates to the exact section — tables included.
02
llmgate — Apache 2.0. One Go client over Anthropic, OpenAI and Gemini, with routing,
fallback and cost tracking on top. LiteLLM, for Go.
03
pdfgrab — MIT. A Go port of pdfplumber: positioned words, font sizes, bounding
boxes and table detection. It is what lets Vectorless read a page at all.
Every product above runs on these three.
Nothing rebuilds them — which is how six
products are the work of one person.
05TrainingThe education arm — and what the last room actually did.
We teach what we build.
Then we publish the register.
176signed up in a fortnight, from one announcement
A final-year medical student,
building what the ward kept needing.
Every product here started as a job I was already doing —
clerking, revising for finals, hunting one paper at one in the morning. I build
in the open and publish what the measurement said — including the
runs that went against me.
A next role, or the backing to launch these. I build either way.
Ten products and nine published libraries, taken from a problem to something
running. I have been the whole team on every one of them — the architecture,
the schema, the interface and the deploy. Either I do that work with a team
around me, or I do it with the funding to put these in front of the people
they were built for. Both interest me; the building happens regardless.